More about HKUST
Crash Reproduction in Android Apps with Stack Trace Only
PhD Thesis Proposal Defence Title: "Crash Reproduction in Android Apps with Stack Trace Only" by Miss Maryam Alsadat MASOUDIAN TARGHI Abstract: As Android applications continue to proliferate, the rising number of reported crashes highlights the critical need for safe and reliable development practices. However, issue tracking systems like GitHub often lack detailed reproduction steps for about 80% of reported cases, with stack traces alone insufficient for effective debugging and crash reproduction. Developers face significant challenges in verifying whether a crash has been resolved due to the vast input space of Android apps, primarily accessed via Graphical User Interfaces (GUIs), and varying environmental settings. This thesis proposes two solutions aimed at narrowing the input space to those that directly contribute to crashes, facilitating effective reproduction of the crashes. The first solution applies a directed fuzzing strategy to concentrate testing on necessary GUI inputs directly tied to crashes. It integrates Attribute-Sensitive Reachability analysis, which simulates the app’s visual state to statically track widget attributes and predict the relevant events from the irrelevant ones leading to crashes before execution. Evaluation on the Themis benchmark shows our directed fuzzing solution reduces crash reproduction time significantly—from six hours down to two hours. The second solution introduces a Neuro-Symbolic approach that combines static program analysis with Large Language Models (LLMs) to infer environmental settings in an Android smartphone that impact the crash occurrences. It relies on the relevancy of API methods’ functionality to environment settings in an Android device to predict possible environment settings affecting a crash occurrence. Using program slicing allows our solution to identify data and control dependencies around crash points, then correlates API functionalities with environment configurations by leveraging LLM-derived specifications. This approach achieves more than 80% recall and precision in detecting relevant settings for 50 crashes collected from highly starred open-sourced Android app repositories in GitHub. Together, our solutions empower developers with effective tools to reproduce crashes more efficiently, ultimately improving the reliability and user experience of Android applications. Date: Tuesday, 3 June 2025 Time: 10:00am - 12:00noon Venue: Room 3494 Lifts 25/26 Committee Members: Prof. Charles Zhang (Supervisor) Dr. Dimitris Papadopoulos (Chairperson) Dr. Shuai Wang